And the award for the most annoying bug to trace, ever, goes to...
which should have been:
The bug was in libneon, and as far I see exists even in their latest vanilla version.
This causes crashes with a frequency which depends on how often TLS connections are being initialized and destroyed. Hours when heavily recycling sessions, or 1-2 minutes if not. (and using tens / hundreds of threads)
There might still be something wrong with the refcounting, but this fix alone seems to make things far more stable. To be continued ( ... but hopefully not )